Confidence and pseudo-log-likelihood
Heavy-only or paired-chain sequence plausibility scoring.
Start this methodScore, represent, restore, and review antibody sequences
Choose an AbLang2 analysis to compare sequence plausibility, create fixed-width representations, restore explicit residue masks, or rank substitutions by model probability.
Choose an approach
Scientific approaches
Select the analysis that best matches your scientific question. Each approach opens with its relevant inputs and controls.
Sequence-level embeddings
One 480-value representation per heavy, light, or paired variable-domain row for clustering and downstream modeling.
Start this methodMasked-residue restoration
Fill literal * masks while retaining chain identity, length, and every unmasked residue.
Start this methodMutation preference ranking
Rank chain-aware substitutions from stepwise AbLang2 token probabilities.
Start this methodPrepare
Antibody variable-domain sequences
Use a protein or annotated-sequence Dataset and map the heavy-chain column, light-chain column, or both. Scoring requires a heavy chain; embeddings, restoration, and mutation preference ranking also accept light-only rows.
- Variable domains should contain 40–180 unambiguous amino acids per supplied chain
- Embeddings: up to 16 rows and 2,880 total residues
- Restoration: up to 8 rows, 1,440 residues, 16 literal * masks per row, and 64 masks total
- Mutation preferences: one row and up to 360 paired residues; scoring: up to 100 rows and 100,000 residues
Heavy-chain column
heavy_sequenceLight-chain column
light_sequenceRestoration mask
EVQLV*SGGGLVQPGGSLRLSC…How the analysis starts
Choose a Project, open a compatible Dataset, then select the rows you want to analyze. Ubi will open this method with the compatible controls and column mappings ready to review.
Configure the scientific method
These controls appear in the Dataset analysis panel, where values can be checked against the actual input before the run starts.
Analysis
Score, embed, restore, or rank mutations
Choose the output needed for the current scientific question; each mode publishes a separate typed Dataset.
Chain mapping
Heavy, light, or paired columns
Map the Dataset fields used for every row; leave one mapping empty for a single-chain analysis.
Scoring depth
Confidence or pseudo-log-likelihood
Use confidence for a fast forward pass or pseudo-log-likelihood for slower residue-masked comparison.
Restoration and mutation controls
Explicit * masks or top 1–100 substitutions
Restoration changes only submitted masks; mutation ranking returns the requested number of model-ranked substitutions.
Review results as scientific outputs
Results open with the figures, structures, sequences, and metrics needed to answer the scientific question. Downloadable files remain available for downstream analysis.
How to interpret the result
- Use scores for relative comparison within a consistently prepared candidate set rather than as an experimental pass/fail threshold.
- Higher model scores correspond to lower pseudo-perplexity; lower pseudo-perplexity means the sequence is more expected by the model.
- Each embedding is one sequence-level 480-value representation, not a residue-level embedding or a measured biological property.
- Restoration evaluates all submitted masks simultaneously, preserves every unmasked residue, and reports probabilities over the 20 canonical amino acids.
- Mutation ranks describe preference under the model sequence distribution; they do not establish improved binding, activity, stability, developability, or immunogenicity.
- Heavy-only, light-only, and paired-chain analyses use different context, so compare like with like unless the study design explicitly accounts for that difference.
1
Candidate score ranking
Compare confidence or pseudo-log-likelihood and pseudo-perplexity across consistently prepared candidates.
2
Sequence representations
Inspect vector summaries in the result view and export complete 480-value embeddings for clustering or downstream models.
3
Restored sequences and mask probabilities
Review each restored residue in sequence context with the complete canonical-amino-acid probability distribution.
4
Chain-aware mutation preferences
Compare source and proposed probabilities, log-probability ratios, one-based positions, and stable ranks.
Continue your analysis
Use a complementary method on the same Project data.
Method scope and limitations
- Scores, vectors, restored residues, and mutation preferences are model-derived evidence and should be reviewed with structural, developability, and assay data.
